sglang 0.5.2rc2__py3-none-any.whl → 0.5.3rc2__py3-none-any.whl

sglang/srt/speculative/ngram_utils.py

@@ -0,0 +1,428 @@
+from __future__ import annotations
+
+import copy
+import logging
+from typing import Optional, Tuple
+
+import torch
+import triton
+
+logger = logging.getLogger(__name__)
+
+from dataclasses import dataclass
+
+import torch.nn.functional as F
+
+from sglang.srt.layers.attention.utils import create_flashinfer_kv_indices_triton
+from sglang.srt.layers.logits_processor import LogitsProcessorOutput
+from sglang.srt.layers.sampler import apply_custom_logit_processor
+from sglang.srt.managers.schedule_batch import (
+    ScheduleBatch,
+    get_last_loc,
+    global_server_args_dict,
+)
+from sglang.srt.sampling.sampling_batch_info import SamplingBatchInfo
+from sglang.srt.speculative.spec_info import SpecInput, SpecInputType
+from sglang.srt.speculative.spec_utils import (
+    TREE_SPEC_KERNEL_AVAILABLE,
+    assign_req_to_token_pool,
+    get_src_tgt_cache_loc,
+    get_target_cache_loc,
+)
+from sglang.srt.utils import is_cuda, is_hip, next_power_of_2
+
+if is_cuda():
+    from sgl_kernel import (
+        top_k_renorm_prob,
+        top_p_renorm_prob,
+        tree_speculative_sampling_target_only,
+        verify_tree_greedy,
+    )
+elif is_hip():
+    from sgl_kernel import verify_tree_greedy
+
+
+@dataclass
+class NgramVerifyInput(SpecInput):
+    def __init__(
+        self,
+        draft_token: torch.Tensor,
+        tree_mask: torch.Tensor,
+        positions: torch.Tensor,
+        retrive_index: torch.Tensor,
+        retrive_next_token: torch.Tensor,
+        retrive_next_sibling: torch.Tensor,
+        draft_token_num: int,
+    ):
+        super().__init__(SpecInputType.NGRAM_VERIFY)
+        self.draft_token = draft_token
+        self.custom_mask = tree_mask
+        self.positions = positions
+        self.retrive_index = retrive_index
+        self.retrive_next_token = retrive_next_token
+        self.retrive_next_sibling = retrive_next_sibling
+        self.draft_token_num = draft_token_num
+        self.device = self.custom_mask.device
+
+    def get_spec_adjust_token_coefficient(self) -> Tuple[int, int]:
+        return self.draft_token_num, self.draft_token_num
+
+    def prepare_for_verify(self, batch: ScheduleBatch, page_size: int):
+        if batch.forward_mode.is_idle():
+            return
+
+        batch.input_ids = self.draft_token
+
+        if page_size == 1:
+            batch.out_cache_loc = batch.alloc_token_slots(len(batch.input_ids))
+            end_offset = batch.seq_lens + self.draft_token_num
+        else:
+            # TODO(lsyin): add prefix lens cpu here to support page size > 1
+            prefix_lens = batch.seq_lens
+            prefix_lens_cpu = batch.seq_lens_cpu
+            end_offset = prefix_lens + self.draft_token_num
+            end_offset_cpu = prefix_lens_cpu + self.draft_token_num
+            last_loc = get_last_loc(
+                batch.req_to_token_pool.req_to_token,
+                batch.req_pool_indices,
+                prefix_lens,
+            )
+            batch.out_cache_loc = batch.alloc_paged_token_slots_extend(
+                prefix_lens,
+                prefix_lens_cpu,
+                end_offset,
+                end_offset_cpu,
+                last_loc,
+                len(batch.input_ids),
+            )
+            self.last_loc = last_loc
+
+        bs = batch.batch_size()
+        assign_req_to_token_pool[(bs,)](
+            batch.req_pool_indices,
+            batch.req_to_token_pool.req_to_token,
+            batch.seq_lens,
+            end_offset,
+            batch.out_cache_loc,
+            batch.req_to_token_pool.req_to_token.shape[1],
+            triton.next_power_of_2(bs),
+        )
+
+    def generate_attn_arg_prefill(
+        self,
+        req_pool_indices: torch.Tensor,
+        paged_kernel_lens: torch.Tensor,
+        paged_kernel_lens_sum: int,
+        req_to_token: torch.Tensor,
+    ):
+        bs = len(req_pool_indices)
+
+        cum_kv_seq_len = torch.zeros((bs + 1,), dtype=torch.int32, device=self.device)
+
+        paged_kernel_lens = paged_kernel_lens + self.draft_token_num
+        cum_kv_seq_len[1:] = torch.cumsum(paged_kernel_lens, dim=0)
+
+        self.qo_indptr = (
+            torch.arange(0, bs + 1, dtype=torch.int32, device=self.device)
+            * self.draft_token_num
+        )
+
+        kv_indices = torch.empty(
+            cum_kv_seq_len[-1], dtype=torch.int32, device=self.device
+        )
+
+        create_flashinfer_kv_indices_triton[(bs,)](
+            req_to_token,
+            req_pool_indices,
+            paged_kernel_lens,
+            cum_kv_seq_len,
+            None,
+            kv_indices,
+            req_to_token.size(1),
+        )
+        return kv_indices, cum_kv_seq_len, self.qo_indptr, self.custom_mask
+
+    def _fill_requests(
+        self,
+        batch: ScheduleBatch,
+        logits_output: torch.Tensor,
+    ):
+        accept_index_cpu = self.accept_index.tolist()
+        predict_cpu = self.predict.tolist()
+        has_finished = False
+
+        # Iterate every accepted token and check if req has finished after append the token
+        # should be checked BEFORE free kv cache slots
+        for i, (req, accept_index_row) in enumerate(zip(batch.reqs, accept_index_cpu)):
+            for j, idx in enumerate(accept_index_row):
+                if idx == -1:
+                    break
+                id = predict_cpu[idx]
+                req.output_ids.append(id)
+                req.check_finished()
+                if req.finished():
+                    has_finished = True
+                    # set all tokens after finished token to -1 and break
+                    self.accept_index[i, j + 1 :] = -1
+                    break
+                else:
+                    if req.grammar is not None:
+                        try:
+                            req.grammar.accept_token(id)
+                        except ValueError as e:
+                            logger.info(
+                                f"{i=}, {req=}\n"
+                                f"{self.accept_index=}\n"
+                                f"{self.predict=}\n"
+                            )
+                            raise e
+            req.spec_verify_ct += 1
+        if has_finished:
+            self.accept_length = (self.accept_index != -1).sum(dim=1) - 1
+        self.accept_index = self.accept_index[self.accept_index != -1]
+
+        logits_output.next_token_logits = logits_output.next_token_logits[
+            self.accept_index
+        ]
+        if logits_output.hidden_states:
+            logits_output.hidden_states = logits_output.hidden_states[self.accept_index]
+        self.verified_id = self.predict[self.accept_index]
+
+    def _free_cache(self, batch: ScheduleBatch, page_size: int):
+        bs = batch.batch_size()
+        # Free the KV cache for unaccepted tokens
+        if page_size == 1:
+            # TODO: boolean array index leads to a device sync. Remove it.
+            evict_mask = torch.full_like(self.draft_token, True, dtype=torch.bool)
+            evict_mask[self.accept_index] = False
+            batch.token_to_kv_pool_allocator.free(batch.out_cache_loc[evict_mask])
+            batch.out_cache_loc = batch.out_cache_loc[self.accept_index]
+        else:
+            # Shift the accepted tokens to the beginning.
+            # Only evict the last part
+            src_cache_loc, tgt_cache_loc, to_free_num_slots = get_src_tgt_cache_loc(
+                batch.seq_lens,
+                batch.out_cache_loc,
+                self.accept_index,
+                self.accept_length,
+                self.draft_token_num,
+                page_size,
+            )
+            to_free_slots = torch.empty(
+                (to_free_num_slots.sum().item(),),
+                dtype=torch.int64,
+                device=to_free_num_slots.device,
+            )
+
+            # out_cache_loc: [0  1  2,  3  4  5,  6  7  8]
+            # accept_index:  [0 -1  2,  3  4 -1,  6 -1 -1]
+            # tgt_cache_loc: [0  1   ,  3  4   ,  6      ]
+            # to_free_slots: [      2,        5,     7  8]
+            # to_free_slots also needs to be page-aligned without the first partial page
+            #
+            # split each row of out_cache_loc into two parts.
+            # 1. the first part goes to tgt_cache_loc. length = accept_length[i] + 1
+            # 2. the second part goes to to_free_slots.
+            get_target_cache_loc[(bs,)](
+                tgt_cache_loc,
+                to_free_slots,
+                self.accept_length,
+                to_free_num_slots,
+                batch.out_cache_loc,
+                self.draft_token_num,
+                next_power_of_2(self.draft_token_num),
+                next_power_of_2(bs),
+            )
+
+            # Free the kv cache
+            batch.token_to_kv_pool_allocator.free(to_free_slots)
+
+            # Copy the kv cache
+            batch.token_to_kv_pool_allocator.get_kvcache().move_kv_cache(
+                tgt_cache_loc, src_cache_loc
+            )
+            batch.out_cache_loc = tgt_cache_loc
+
+        assign_req_to_token_pool[(bs,)](
+            batch.req_pool_indices,
+            batch.req_to_token_pool.req_to_token,
+            batch.seq_lens,
+            batch.seq_lens + self.accept_length + 1,
+            batch.out_cache_loc,
+            batch.req_to_token_pool.req_to_token.shape[1],
+            triton.next_power_of_2(bs),
+        )
+
+    def _greedy_verify(
+        self,
+        batch: ScheduleBatch,
+        logits_output: LogitsProcessorOutput,
+    ):
+        bs = batch.batch_size()
+        target_predict = torch.argmax(logits_output.next_token_logits, dim=-1)
+        target_predict = target_predict.reshape(bs, self.draft_token_num)
+
+        candidates = self.draft_token.reshape(bs, self.draft_token_num)
+        predict_shape = list(logits_output.next_token_logits.shape)[:-1]
+        predict_shape[-1] += 1
+        self.predict = torch.empty(predict_shape, dtype=torch.int32, device=self.device)
+        self.accept_index = torch.full(
+            (bs, self.draft_token_num), -1, dtype=torch.int32, device=self.device
+        )
+        self.accept_length = torch.empty((bs,), dtype=torch.int32, device=self.device)
+
+        verify_tree_greedy(
+            predicts=self.predict,  # mutable
+            accept_index=self.accept_index,  # mutable
+            accept_token_num=self.accept_length,  # mutable
+            candidates=candidates,
+            retrive_index=self.retrive_index,
+            retrive_next_token=self.retrive_next_token,
+            retrive_next_sibling=self.retrive_next_sibling,
+            target_predict=target_predict,
+        )
+
+    def _sampling_verify(
+        self,
+        batch: ScheduleBatch,
+        logits_output: LogitsProcessorOutput,
+        sampling_info: SamplingBatchInfo,
+    ):
+        bs = batch.batch_size()
+        candidates = self.draft_token.reshape(bs, self.draft_token_num)
+        predict_shape = list(logits_output.next_token_logits.shape)[:-1]
+        predict_shape[-1] += 1
+        self.predict = torch.empty(predict_shape, dtype=torch.int32, device=self.device)
+        self.accept_index = torch.full(
+            (bs, self.draft_token_num), -1, dtype=torch.int32, device=self.device
+        )
+        self.accept_length = torch.empty((bs,), dtype=torch.int32, device=self.device)
+        # apply temperature and get target probs
+        expanded_temperature = torch.repeat_interleave(
+            sampling_info.temperatures, self.draft_token_num, dim=0
+        )  # (bs * draft_token_num, 1)
+
+        target_probs = F.softmax(
+            logits_output.next_token_logits / expanded_temperature, dim=-1
+        )  # (bs * draft_token_num, vocab_size)
+
+        # NOTE: The test shows that top_p_renorm_prob and top_k_renorm_prob are the key factors
+        # contributing to the poor performance of _sampling_verify.
+        target_probs = top_k_renorm_prob(
+            target_probs,
+            torch.repeat_interleave(sampling_info.top_ks, self.draft_token_num, dim=0),
+        )  # (bs * draft_token_num, vocab_size)
+
+        if sampling_info.need_top_p_sampling:
+            # logger.info("Using top-p sampling in speculative decoding verification.")
+            target_probs = top_p_renorm_prob(
+                target_probs,
+                torch.repeat_interleave(
+                    sampling_info.top_ps, self.draft_token_num, dim=0
+                ),
+            )
+
+        target_probs = target_probs.reshape(bs, self.draft_token_num, -1)
+        draft_probs = torch.zeros(
+            target_probs.shape, dtype=torch.float32, device=self.device
+        )
+
+        # coins for rejection sampling
+        coins = torch.rand_like(candidates, dtype=torch.float32, device=self.device)
+        # coins for final sampling
+        coins_for_final_sampling = torch.rand(
+            (bs,), dtype=torch.float32, device=self.device
+        )
+        tree_speculative_sampling_target_only(
+            predicts=self.predict,  # mutable
+            accept_index=self.accept_index,  # mutable
+            accept_token_num=self.accept_length,  # mutable
+            candidates=candidates.to(torch.int64),
+            retrive_index=self.retrive_index.to(torch.int64),
+            retrive_next_token=self.retrive_next_token.to(torch.int64),
+            retrive_next_sibling=self.retrive_next_sibling.to(torch.int64),
+            uniform_samples=coins,
+            uniform_samples_for_final_sampling=coins_for_final_sampling,
+            target_probs=target_probs,
+            draft_probs=draft_probs,
+            threshold_single=global_server_args_dict[
+                "speculative_accept_threshold_single"
+            ],
+            threshold_acc=global_server_args_dict["speculative_accept_threshold_acc"],
+            deterministic=True,
+        )
+
+    def verify(
+        self,
+        batch: ScheduleBatch,
+        logits_output: LogitsProcessorOutput,
+        page_size: int,
+        vocab_mask: Optional[torch.Tensor] = None,  # For grammar
+    ) -> torch.Tensor:
+        bs = self.retrive_index.shape[0]
+        sampling_info = batch.sampling_info
+
+        if bs != len(sampling_info):
+            sampling_info = copy.deepcopy(sampling_info)
+            # NOTE: retrive_index are the indices of the requests that are kept.
+            sampling_info.filter_batch(self.retrive_index.tolist(), self.retrive_index)
+
+        # Apply the custom logit processors if registered in the sampling info.
+        if sampling_info.has_custom_logit_processor:
+            apply_custom_logit_processor(
+                logits_output.next_token_logits,
+                sampling_info,
+                num_tokens_in_batch=self.draft_token_num,
+            )
+
+        # Apply penalty
+        if sampling_info.penalizer_orchestrator.is_required:
+            # This is a relaxed version of penalties for speculative decoding.
+            linear_penalty = torch.zeros(
+                (bs, logits_output.next_token_logits.shape[1]),
+                dtype=torch.float32,
+                device=self.device,
+            )
+            sampling_info.apply_logits_bias(linear_penalty)
+            logits_output.next_token_logits.add_(
+                torch.repeat_interleave(linear_penalty, self.draft_token_num, dim=0)
+            )
+
+        # Apply grammar mask
+        if vocab_mask is not None:
+            assert self.grammar is not None
+            self.grammar.apply_vocab_mask(
+                logits=logits_output.next_token_logits, vocab_mask=vocab_mask
+            )
+
+        # Sample tokens. Force greedy sampling on AMD
+        is_all_greedy = sampling_info.is_all_greedy
+        if (not is_all_greedy) and (not TREE_SPEC_KERNEL_AVAILABLE):
+            logger.warning(
+                "Tree speculative sampling kernel unavailable (likely AMD/HIP build). "
+                "Falling back to greedy verification."
+            )
+
+        if is_all_greedy or not TREE_SPEC_KERNEL_AVAILABLE:
+            self._greedy_verify(batch, logits_output)
+        else:
+            # NOTE: Compared with greedy_verify, the performance of _sampling_verify is relatively poor.
+            self._greedy_verify(batch, logits_output)
+            # self._sampling_verify(batch, logits_output, sampling_info)
+
+        self._fill_requests(batch, logits_output)
+        self._free_cache(batch, page_size)
+
+        accept_length_cpu = self.accept_length.cpu()
+        num_accepted_tokens = accept_length_cpu.sum().item()
+
+        batch.seq_lens.add_(self.accept_length + 1)
+        batch.seq_lens_cpu.add_(accept_length_cpu + 1)
+
+        return logits_output, self.verified_id, num_accepted_tokens
+
+    def filter_batch(self, new_indices: torch.Tensor, has_been_filtered: bool = True):
+        pass
+
+    def merge_batch(self, spec_info: NgramVerifyInput):
+        pass

sglang/srt/speculative/ngram_worker.py

@@ -0,0 +1,245 @@
+import logging
+from typing import List, Optional
+
+import numpy as np
+import torch
+from sgl_kernel.speculative import reconstruct_indices_from_tree_mask
+
+from sglang.srt.managers.schedule_batch import ScheduleBatch
+from sglang.srt.managers.tp_worker import TpModelWorker
+from sglang.srt.model_executor.forward_batch_info import ForwardBatchOutput, ForwardMode
+from sglang.srt.server_args import ServerArgs
+from sglang.srt.speculative.cpp_ngram.ngram_cache import NgramCache
+from sglang.srt.speculative.ngram_utils import NgramVerifyInput
+from sglang.srt.speculative.spec_info import SpeculativeAlgorithm
+
+logger = logging.getLogger(__name__)
+
+USE_FULL_MASK = True
+
+
+class NGRAMWorker:
+    def __init__(
+        self,
+        server_args: ServerArgs,
+        gpu_id: int,
+        tp_rank: int,
+        dp_rank: Optional[int],
+        moe_ep_rank: int,
+        nccl_port: int,
+        target_worker: TpModelWorker,
+    ):
+        self.target_worker = target_worker
+        self.model_runner = target_worker.model_runner
+        self.tp_rank = tp_rank
+        self.page_size = server_args.page_size
+        self.draft_token_num: int = server_args.speculative_num_draft_tokens
+        self.branch_length: int = server_args.speculative_ngram_branch_length
+        self.max_match_window_size: int = (
+            server_args.speculative_ngram_max_match_window_size
+        )
+
+        self.max_batch_size = target_worker.max_running_requests
+        self.device = f"cuda:{gpu_id}" if gpu_id >= 0 else "cuda"
+
+        self._init_preallocated_tensors()
+
+        self.ngram_cache = NgramCache(
+            min_match_window_size=server_args.speculative_ngram_min_match_window_size,
+            max_match_window_size=server_args.speculative_ngram_max_match_window_size,
+            min_bfs_breadth=server_args.speculative_ngram_min_bfs_breadth,
+            max_bfs_breadth=server_args.speculative_ngram_max_bfs_breadth,
+            capacity=server_args.speculative_ngram_capacity,
+            branch_length=server_args.speculative_ngram_branch_length,
+            draft_token_num=server_args.speculative_num_draft_tokens,
+        )
+
+    def clear_cache_pool(self):
+        self.ngram_cache.reset()
+
+    def _efficient_concat_last_n(self, seq1: List[int], seq2: List[int], n: int):
+        seq2_len = len(seq2)
+        if seq2_len >= n:
+            return seq2[-n:]
+
+        need_from_seq1 = n - seq2_len
+        return seq1[-need_from_seq1:] + seq2
+
+    def _init_preallocated_tensors(self):
+        max_total_drafts = self.max_batch_size * self.draft_token_num
+        max_total_mask_size = (
+            self.max_batch_size * self.draft_token_num * self.draft_token_num
+        )
+
+        self.draft_tokens = torch.empty(
+            (max_total_drafts,), dtype=torch.int64, device=self.device
+        )
+        self.retrieve_indexes = torch.empty(
+            (self.max_batch_size, self.draft_token_num),
+            dtype=torch.int64,
+            device=self.device,
+        )
+        self.retrive_next_token = torch.empty(
+            (self.max_batch_size, self.draft_token_num),
+            dtype=torch.int64,
+            device=self.device,
+        )
+        self.retrive_next_sibling = torch.empty(
+            (self.max_batch_size, self.draft_token_num),
+            dtype=torch.int64,
+            device=self.device,
+        )
+        self.positions = torch.empty(
+            (max_total_drafts,), dtype=torch.int64, device=self.device
+        )
+        self.tree_mask = torch.empty(
+            (max_total_mask_size,), dtype=torch.bool, device=self.device
+        )
+
+        self.draft_tokens_batch = []
+        self.tree_mask_batch = []
+        self.retrieve_indexes_batch = []
+        self.retrive_next_token_batch = []
+        self.retrive_next_sibling_batch = []
+        self.positions_batch = []
+
+        for bs in range(0, self.max_batch_size + 1):
+            self.retrieve_indexes_batch.append(self.retrieve_indexes[:bs, :])
+            self.retrive_next_token_batch.append(self.retrive_next_token[:bs, :])
+            self.retrive_next_sibling_batch.append(self.retrive_next_sibling[:bs, :])
+            self.positions_batch.append(self.positions[: bs * self.draft_token_num])
+            self.draft_tokens_batch.append(
+                self.draft_tokens[: bs * self.draft_token_num]
+            )
+            self.tree_mask_batch.append(
+                self.tree_mask[: bs * self.draft_token_num * self.draft_token_num]
+            )
+
+    def _prepare_draft_tokens(
+        self, batch: ScheduleBatch
+    ) -> tuple[np.ndarray, np.ndarray]:
+        bs = batch.batch_size()
+
+        self.ngram_cache.synchronize()
+        batch_tokens = []
+        for req in batch.reqs:
+            check_token = self._efficient_concat_last_n(
+                req.origin_input_ids, req.output_ids, self.max_match_window_size
+            )
+            batch_tokens.append(check_token)
+        req_drafts, mask = self.ngram_cache.batch_get(batch_tokens)
+        total_draft_token_num = len(req_drafts)
+
+        # Check if speculative decoding is needed; here we always enforce it
+        assert (
+            total_draft_token_num == bs * self.draft_token_num
+        ), f"{total_draft_token_num=}, {bs=}, {self.draft_token_num=}"
+        return req_drafts, mask
+
+    def _prepare_for_speculative_decoding(self, batch: ScheduleBatch):
+        if batch.forward_mode.is_extend():
+            return
+
+        bs = batch.batch_size()
+
+        retrive_index = self.retrieve_indexes_batch[bs]
+        retrive_next_token = self.retrive_next_token_batch[bs]
+        retrive_next_sibling = self.retrive_next_sibling_batch[bs]
+        positions = self.positions_batch[bs]
+        tree_mask = self.tree_mask_batch[bs]
+        draft_tokens = self.draft_tokens_batch[bs]
+
+        req_drafts, mask = self._prepare_draft_tokens(batch)
+        tree_mask.copy_(torch.from_numpy(mask), non_blocking=True)
+        draft_tokens.copy_(torch.from_numpy(req_drafts), non_blocking=True)
+
+        reconstruct_indices_from_tree_mask(
+            tree_mask,
+            batch.seq_lens,
+            positions,  # mutable
+            retrive_index,  # mutable
+            retrive_next_token,  # mutable
+            retrive_next_sibling,  # mutable
+            bs,
+            self.draft_token_num,
+        )
+
+        # NOTE: QLEN_MASK is faster than FULL_MASK, but requires corresponding changes in flashinfer.
+        # Testing shows about 8% performance improvement (the effect is roughly proportional to batch size).
+        if USE_FULL_MASK:
+            tree_mask = []
+            mask = mask.reshape(
+                batch.batch_size(), self.draft_token_num, self.draft_token_num
+            )
+            for i, req in enumerate(batch.reqs):
+                seq_len = len(req.origin_input_ids) + len(req.output_ids)
+                req_mask = torch.ones((self.draft_token_num, seq_len - 1)).cuda()
+                req_mask = torch.cat(
+                    (req_mask, torch.from_numpy(mask[i]).cuda()), dim=1
+                ).to(torch.bool)
+                tree_mask.append(req_mask.flatten())
+            tree_mask = torch.cat(tree_mask, dim=0)
+
+        batch.spec_algorithm = SpeculativeAlgorithm.NGRAM
+        batch.forward_mode = ForwardMode.TARGET_VERIFY
+        batch.spec_info = NgramVerifyInput(
+            draft_tokens,
+            tree_mask,
+            positions,
+            retrive_index,
+            retrive_next_token,
+            retrive_next_sibling,
+            self.draft_token_num,
+        )
+        batch.spec_info.prepare_for_verify(batch, self.page_size)
+
+    def _update_ngram_cache(self, batch: ScheduleBatch):
+        batch_tokens = []
+        for req in batch.reqs:
+            # FIXME: Whether to insert 'extend' into the cache or not, after testing,
+            # there is not much difference, so we will not insert it for now.
+            # if batch.forward_mode.is_extend():
+            #     put_ids = req.origin_input_ids + req.output_ids
+            # else:
+            put_ids = self._efficient_concat_last_n(
+                req.origin_input_ids, req.output_ids, self.branch_length
+            )
+            batch_tokens.append(put_ids)
+        self.ngram_cache.batch_put(batch_tokens)
+
+    def forward_batch_generation(self, batch: ScheduleBatch) -> ForwardBatchOutput:
+        self._prepare_for_speculative_decoding(batch)
+        model_worker_batch = batch.get_model_worker_batch()
+        num_accepted_tokens = 0
+
+        if model_worker_batch.forward_mode.is_target_verify():
+            forward_batch_output = self.target_worker.forward_batch_generation(
+                model_worker_batch, is_verify=True
+            )
+            logits_output, can_run_cuda_graph = (
+                forward_batch_output.logits_output,
+                forward_batch_output.can_run_cuda_graph,
+            )
+            verify_input = model_worker_batch.spec_info
+            logits_output, next_token_ids, num_accepted_tokens = verify_input.verify(
+                batch, logits_output, self.page_size
+            )
+            self._update_ngram_cache(batch)
+            batch.forward_mode = ForwardMode.DECODE
+
+        else:
+            forward_batch_output = self.target_worker.forward_batch_generation(
+                model_worker_batch
+            )
+            logits_output, next_token_ids, can_run_cuda_graph = (
+                forward_batch_output.logits_output,
+                forward_batch_output.next_token_ids,
+                forward_batch_output.can_run_cuda_graph,
+            )
+
+        return ForwardBatchOutput(
+            logits_output=logits_output,
+            next_token_ids=next_token_ids,
+            num_accepted_tokens=num_accepted_tokens,
+            can_run_cuda_graph=can_run_cuda_graph,
+        )